Relay
Relay takes legacy accounting data through migration QA to a verified ERP go-live: it validates and
reconciles the data deterministically, turns failures into a ranked work queue, and gates go-live
on a reproducible run.
Live demo: https://web-production-9032d5.up.railway.app — opens on the command center, no
login, view-only. Start with Open implementation, then the work queue.
Result on the demo migration: 37 rules and 10 reconciliations over ~24,000 records raise 65
findings, grouped into the 20 decisions a person has to make, blocking work first, with
$217,212.85 of unresolved exposure traced to source lines.
- Ingest legacy ledgers, subledgers, control reports and bank data; source rows are immutable and
keep their file and line.
- Process: map onto a canonical model, then run 37 rules and 10 reconciliations against the
customer's own control reports.
- Operate: a work queue of decisions, a bounded read-only investigator that gathers evidence and
drafts a fix, and change requests with segregation-of-duties approval.
- Output: 12 readiness gates on a fingerprinted run, re-run after every fix, with a
hash-chained audit log.
Stack: Python 3.12, FastAPI, SQLAlchemy 2.0, PostgreSQL 16 · Next.js 16, React 19, TypeScript ·
Docker · Playwright · deployed on Railway.
make setup && make up && make demo-reset # then open http://127.0.0.1:3000
Demo runs on fictional data; its AI investigator is a scripted replay, not a model.
The problem
An ERP implementation team has to move messy legacy accounting data into a new system and then
answer one question: is this migration actually safe to launch? Today that answer is assembled
by hand — spreadsheets comparing exports against trial balances, mapping reviews in email, fixes
applied straight to staging data, and a go-live decision that comes down to "looks good" in a
meeting.
What Relay does
Relay takes legacy accounting data through deterministic validation, reconciliation,
investigation, governed remediation and verified launch readiness:
Import → Profile → Map → Normalize → Validate → Reconcile → Investigate → Resolve / Approve →
Verify → Launch
- Imports legacy ledgers, subledgers, control reports and bank data. Source rows are immutable
and keep their file and line; unreadable rows are quarantined, never guessed at.
- Maps and normalizes columns and legacy accounts onto a canonical model and a target chart of
accounts, through versioned mapping sets.
- Validates and reconciles with 37 deterministic rules and 10 reconciliations against the
customer's own control reports. Every discrepancy drills down to the records and source lines
behind it.
- Turns failures into a work queue: 65 findings become the 20 decisions a person actually has
to make, blocking work first, each with the money attached.
- Investigates with a bounded investigator that traces an issue back to evidence and drafts a
fix.
- Governs every correction as a change request that people other than the requester must
approve, then re-runs the deterministic checks to prove the fix worked.
- Decides go-live with 12 readiness gates on a reproducible run, and records all of it in an
append-only, hash-chained audit log.
The design principle
Deterministic software detects and verifies financial failures. The bounded investigator does the
expensive evidence gathering and proposes a resolution. Humans keep authority over every material
change. Then the system re-runs its deterministic controls to verify the approved fix.
Turning the AI off removes no correctness — the end-to-end suite runs the whole product with it
disabled.
Why I built it
Data migration is where ERP implementations are won or lost, and most of the work is repeatable:
comparing exports, chasing differences, proving a fix held. I wanted to see how much of that loop
can be automated without handing judgement or authority to a model — software doing the
checking and the evidence gathering, people making the calls, and every number traceable back to a
line in a file.
A tour
The investigation page is where the whole loop becomes visible on one screen: the deterministic
finding it started from, every read-only tool call, the investigator's reading of that evidence
kept visually separate from it, Relay's verdict on each claim, and the decision left to a person.
More screens — the work-item detail, the record inspector, a change request through its whole
lifecycle, and the audit log — are in docs/images/. A three-minute walkthrough of
the live demo is in docs/demo-script.md.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
